DECISION RULES EXTRACTION FROM NEURAL NETWORK: A MODIFIED PEDAGOGICAL APPROACH

Abstract

Neural network (NN) methods are sometimes useless in practical applications, because they are not pro-perly tailored to the particular market’s needs. We focus thereinafter specifically on financial market applications. NNs have not gained full acceptance here yet. One of the main reasons is the “Black Box” problem (lack of the NN deci-sions explanatory power). It remains an open issue for the top and middle managerial level. There are though some NN decisions rule extraction methods like decompositional, pedagogical or eclectic, but they suffer from low portability of the rule extraction technique across various neural net architectures, high level of granularity, algorithmic sophisti-cation of the rule extraction technique etc. The author propose to eliminate some known drawbacks using an innovative extension of the pedagogical approach. The idea is exposed by the use of a widespread MLP neural net (as a common tool for the decisions’ space fragmentation in the financial problems’ domain) and SOM (for the input data space clus-terization). The feedback of both nets’ performance is related and targeted through the iteration cycle by achievement of the best matching between the decision space fragments and input data space clusters. Three sets of rules are generated algorithmically or by fuzzy membership functions. Empirical validation of the common financial benchmark problems is conducted with an appropriately prepared software solution.